AutoPersonas: A Multi-Timescale Loop Engine for Open-Ended Persona Evolution

🕒 Published (v1): 2026-07-09 08:56 UTC · Source: Arxiv · link

Why this paper was selected

Novel self-locking failure mode diagnosis in long-running persona agents; multi-timescale loop engine

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AutoPersonas is a multi-timescale life-environment engine for long-term LLM persona agents that tackles "self-locking"—the runtime collapse where a continuing persona loops back to familiar environments, stale life stages, and unresolved decisions. It separates environment-side Occurrences, accumulated Observations, and persona State (the OSO loop), then applies context-slice masking and divergence targeting to break context gravity. Diagnostic simulations across eight models confirm the anti-fixation regime substantially reduces macro-theme repetition while preserving identity continuity.

Problem

Long-horizon LLM persona agents suffer from self-locking: even when generating locally plausible text, the life trajectory collapses—returning to the same places, unresolved choices, relationships, and life stages. Two coupled pressures cause this: (1) model-level convergence toward high-probability behavioral channels, and (2) system-level context gravity, where State, memory, and environment summaries continuously pull future generation back to prior attractors. Existing approaches (more memory, random novelty injection, sandbox society simulators like Generative Agents/Agentopia) do not solve this because they either amplify stale state or break identity continuity.

Method

AutoPersonas introduces a three-layer separation: macro-world (shared backdrop), life-environment (persona-specific dynamic radius of places, people, routines, opportunities), and persona layer (identity, values, State). The core loop is the OSO loop:

1. Occurrences — environment-side future-facing material generated with context-slice masking (withholding prior State from the divergence source) and per-sample divergence targeting (conditioning on withheld, persona-conditioned targets) to prevent past State from dominating.
2. Observations — Occurrences must survive evidence review before entering the system; they are logged as candidate evidence, not immediately absorbed into State.
3. State — a semantic State machine: schema-bounded dimensions with LLM semantic integration to detect when sparse accumulated Observations justify State-level revision, handling both slow drift and sudden jumps without a handcrafted ontology.

A key empirical design finding: when historical/current-state content occupies more than ~60% of the available token budget in future-facing generation calls, divergence degrades. This motivates the context-governance architecture.

Key Contributions

• Defines self-locking as a runtime functional trajectory collapse distinct from model collapse or output-diversity degradation.
• Formalizes the macro-world / life-environment / persona three-layer distinction, arguing open-ended growth requires a persona-conditioned life-environment layer rather than a shared world simulator.
• Introduces the OSO loop as a transmission mechanism enforcing evidence-governed absorption before State changes propagate.
• Proposes the semantic State machine: LLM-evaluated evidence accumulation inside schema-bounded State dimensions, without enumerated state transitions.
• Introduces day-level simulation as the appropriate audit granularity; coarser (weekly/yearly) summaries mask micro-failures and recursive indecision.
• Provides a diagnostic audit suite: environment watermark shells, occurrence-hardening gaps, slow-change accumulation failures, recursive indecision, relationship persistence—plus an eight-model quantitative action-category repetition stress test.
• Demonstrates that context-slice masking + divergence targeting constitutes a practical bounded anti-fixation intervention.
• Frames AutoPersonas as part of SoulOS, a broader three-system architecture; this paper's scope is the life-environment engine only.

Results

• Baseline self-locking (action-channel stress test): Eight models, 40-day action-only direct loop. Mean rolling 5-day action-category repetition: 95.2%–97.6%; all models crossed 80% repetition by day 9 and 90% by day 11.
• Semantic re-keying: Same 40-day outputs, macro-theme repetition 79.0%–88.0% across all eight models.
• A/B intervention (same runtime, 40 days): Context-slice masking + per-sample divergence targeting reduced macro-theme repetition from 61.8% → 36.3% in the masked lane (vs. unmasked baseline); cumulative theme count roughly doubled from 55 → 102.
• Fictional-world generalization: Juvenile-goblin setting reproduced the anti-fixation regime at 42.8% blended repetition with no hard real-world intrusions.
• Three-year compressed simulation: Exposed environment watermark shells, occurrence-hardening gaps, slow-change accumulation failures, recursive indecision, and weak relationship persistence (qualitative audit).

Limitations

• No standardized benchmark for open-ended persona evolution; evaluations rely on researcher diagnostic judgment rather than external ground truth.
• The ~60% context-budget threshold for divergence degradation is a design heuristic from repeated researcher observation, not a controlled empirical law.
• Self-locking audits require day-level granularity; the cost of running long compressed simulations is high.
• The paper does not evaluate the full SoulOS stack (CIBE dialogue layer, dual-recall at response time) — only the life-environment engine is assessed.
• Social feedback in all experiments comes from AI model outputs, not humans; real human interaction dynamics are untested.
• Relationship persistence remains a noted weakness even after the anti-fixation intervention (flagged in diagnostic audit but not quantitatively resolved).

Relevance to Agentic AI / LLM Agents

AutoPersonas addresses a foundational problem for persistent, always-on LLM agents: context gravity and recursive state authority cause long-running agents to degrade into behavioral loops regardless of model capability. The OSO loop and semantic State machine are directly applicable to any architecture where an agent must accumulate evidence and revise its own operating state over extended time horizons—not just companion personas but also autonomous agents that self-manage context between tasks. The distinction between macro-world simulation and persona-conditioned life-environment is a useful architectural contribution for multi-agent systems where agents should inhabit meaningfully different effective environments even when sharing a world model. The diagnostic audit methodology (action-channel repetition, macro-theme diversity, watermark shells) provides a concrete evaluation framework that the broader agentic AI field currently lacks for long-horizon open-ended behavior.